andrea.pedrotti
/
gFun
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Implemented funnelling architecture

This commit is contained in:
andrea 2021-01-26 10:15:55 +01:00
parent a5af2134bf
commit 108f423d41
3 changed files with 18 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
refactor/util/SIF_embed.py
View File

@ -1,6 +1,7 @@
import numpy as np import numpy as np
from sklearn.decomposition import TruncatedSVD from sklearn.decomposition import TruncatedSVD
def get_weighted_average(We, x, w): def get_weighted_average(We, x, w):
""" """
Compute the weighted average vectors Compute the weighted average vectors
@ -15,6 +16,7 @@ def get_weighted_average(We, x, w):
emb[i,:] = w[i,:].dot(We[x[i,:],:]) / np.count_nonzero(w[i,:]) emb[i,:] = w[i,:].dot(We[x[i,:],:]) / np.count_nonzero(w[i,:])
return emb return emb
def compute_pc(X,npc=1): def compute_pc(X,npc=1):
""" """
Compute the principal components. Compute the principal components.
@ -26,6 +28,7 @@ def compute_pc(X,npc=1):
svd.fit(X) svd.fit(X)
return svd.components_ return svd.components_
def remove_pc(X, npc=1): def remove_pc(X, npc=1):
""" """
Remove the projection on the principal components Remove the projection on the principal components
@ -34,7 +37,7 @@ def remove_pc(X, npc=1):
:return: XX[i, :] is the data point after removing its projection :return: XX[i, :] is the data point after removing its projection
""" """
pc = compute_pc(X, npc) pc = compute_pc(X, npc)
if npc==1: if npc == 1:
XX = X - X.dot(pc.transpose()) * pc XX = X - X.dot(pc.transpose()) * pc
else: else:
XX = X - X.dot(pc.transpose()).dot(pc) XX = X - X.dot(pc.transpose()).dot(pc)

2
refactor/util/common.py
View File

@ -368,4 +368,4 @@ def get_params(optimc=False):
return None return None
c_range = [1e4, 1e3, 1e2, 1e1, 1, 1e-1] c_range = [1e4, 1e3, 1e2, 1e1, 1, 1e-1]
kernel = 'rbf' kernel = 'rbf'
return [{'kernel': [kernel], 'C': c_range, 'gamma':['auto']}] return [{'kernel': [kernel], 'C': c_range, 'gamma':['auto']}]

15
refactor/view_generators.py
View File

@ -41,17 +41,20 @@ class ViewGen(ABC):
class VanillaFunGen(ViewGen): class VanillaFunGen(ViewGen):
def __init__(self, base_learner, n_jobs=-1): def __init__(self, base_learner, first_tier_parameters=None, n_jobs=-1):
""" """
Original funnelling architecture proposed by Moreo, Esuli and Sebastiani in DOI: https://doi.org/10.1145/3326065 Original funnelling architecture proposed by Moreo, Esuli and Sebastiani in DOI: https://doi.org/10.1145/3326065
:param base_learner: naive monolingual learners to be deployed as first-tier learners. Should be able to :param base_learner: naive monolingual learners to be deployed as first-tier learners. Should be able to
return posterior probabilities. return posterior probabilities.
:param base_learner:
:param n_jobs: integer, number of concurrent workers :param n_jobs: integer, number of concurrent workers
""" """
super().__init__() super().__init__()
self.learners = base_learner self.learners = base_learner
self.first_tier_parameters = first_tier_parameters
self.n_jobs = n_jobs self.n_jobs = n_jobs
self.doc_projector = NaivePolylingualClassifier(self.learners) self.doc_projector = NaivePolylingualClassifier(base_learner=self.learners,
parameters=self.first_tier_parameters, n_jobs=self.n_jobs)
self.vectorizer = TfidfVectorizerMultilingual(sublinear_tf=True, use_idf=True) self.vectorizer = TfidfVectorizerMultilingual(sublinear_tf=True, use_idf=True)
def fit(self, lX, lY): def fit(self, lX, lY):
@ -61,8 +64,16 @@ class VanillaFunGen(ViewGen):
return self return self
def transform(self, lX): def transform(self, lX):
"""
(1) Vectorize documents
(2) Project them according to the learners SVMs
(3) Apply L2 normalization to the projection
:param lX:
:return:
"""
lX = self.vectorizer.transform(lX) lX = self.vectorizer.transform(lX)
lZ = self.doc_projector.predict_proba(lX) lZ = self.doc_projector.predict_proba(lX)
lZ = _normalize(lZ, l2=True)
return lZ return lZ
def fit_transform(self, lX, ly): def fit_transform(self, lX, ly):